MOBILE AND UBIQUITOUS SYSTEMS

www.computer.org/pervasive

Cellular Census: Explorations in Urban Data

Collection
Jonathan Reades, Francesco Calabrese, Andres Sevtsuk, and Carlo Ratti

Vol. 6, No. 3
July–September 2007

This material is presented to ensure timely dissemination of scholarly and technical

work. Copyright and all rights therein are retained by authors or by other copyright
holders. All persons copying this information are expected to adhere to the terms
and constraints invoked by each author's copyright. In most cases, these works
may not be reposted without the explicit permission of the copyright holder.

© 2007 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or

for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be

obtained from the IEEE.

For more information, please see www.ieee.org/web/publications/rights/index.html.


Cellular Census:
Explorations in Urban
Data Collection
Analysis of cell phone use can provide an important new way of looking
at the city as a holistic, dynamic system.
M
uch of our understanding of ur-
ban systems comes from tradi-
tional data collection methods
such as surveys by person or
phone. These approaches can
provide detailed information about urban behav-
iors, but they’re hard to update and might limit
results to “snapshots in time.”
In the past few years, some innovative ap-
proaches have sought to use mobile devices to col-
lect spatiotemporal data (see the sidebar, “Urban
Analysis Using Mobile-Device Data”). But little
research has been done to develop and analyze the
much larger samples of existing data generated daily
by mobile networks.
The most common explanation for this is that
the challenge of data-sharing with the telecom-
munications industry has ham-
Jonathan Reades pered data access. However, in
University College London early 2006, a collaboration be-
tween Telecom Italia, which
Francesco Calabrese, Andres
serves 40 percent of the Roman
Sevtsuk, and Carlo Ratti
market, and MIT’s SENSEable
Massachusetts Institute City Laboratory (http://senseable.
of Technology mit.edu) allowed unprecedented
access to aggregate mobile phone
data from Rome. Here, we ex-
plore how researchers might be able to use data
for an entire metropolitan region to analyze
urban dynamics.
The Real Time Rome platform
The TI and MIT collaboration, developed
under the Real Time Rome label, was shown at
the 2006 Venice Biennale. The installation incor-
porated both real-time and historical visualiza-
30 PERVASIVE computing Published by the IEEE Computer Society ■ 1536-1268/07/$25.00 © 2007 IEEE
URBAN COMPUTING
tions of mobile phone usage levels in central
Rome during autumn 2006. The system archi-
tecture, including data collection, transfer, and
processing, has been detailed elsewhere.1
TI supplied several different types of data, first
and foremost of which was the Erlang, a mea-
sure of network bandwidth usage typically col-
lected at the antenna level. Additionally, TI used
its innovative Lochness platform to supply aggre-
gate location and trajectory data on callers using
the system for more than three minutes at a time.
Two transportation companies—Atac-Rome (a
public bus company) and Samarcanda (a private
taxi company)—also provided supplemental GPS
data to MIT for further processing. However,
here we focus on the Erlang data collected over
four months in late 2006 and covering a region
of 47 km2, considering how it can help us better
understand urban dynamics.
An Erlang is one person-hour of phone use, so
1 Erlang could represent one person talking for an
hour, two people talking for a half hour each, 30
people speaking for two minutes each, and so on.
Consequently, Erlang data is both aggregate and
anonymous, and deducing individual identities
from the data collected and stored in the system is
impossible. Additionally, because Erlang data is a
standard measure used by most network opera-
tors, it’s an accessible source for the analysis of typ-
ical GSM (Global System for Mobile Communi-
cation) networks. You can collect Erlang data
without installing new applications or upgrading
the base station controllers, both of which incur
costs and operational risks for the networks.
Although Erlang data can’t be linked to an indi-
vidual subscriber and doesn’t offer the locational
 Urban Analysis Using Mobile-Device Data

T he Massachusetts Institute of Technology’s Reality Mining project

(http://reality.media.mit.edu) successfully abstracted common
behavioral patterns from the activities of 70 students and faculty issued
quire user consent. Consequently, sample sizes are necessarily more
modest and might be limited in terms of the research’s spatial extent.

REFERENCES
with Nokia phones carrying specially designed logging software.1,2
Rein Ahas and Ülar Mark tracked the mobile phones of 300 users 1. N. Eagle and A. Pentland, “Eigenbehaviors: Identifying Structure in Rou-
for a “social positioning method” analysis.3 By combining spatio- tine,” 2006; http://vismod.media.mit.edu//tech-reports/TR-601.pdf.

temporal data from phones with demographic and attitudinal data 2. N. Eagle and A. Pentland, “Reality Mining: Sensing Complex Social Sys-
from surveys, they created a map of social spaces in Estonia. tems,” Personal and Ubiquitous Computing, vol. 10, no. 4, 2006, pp.
In the UK, the Cityware research group has taken a more readily 255–268.

scalable approach. They supplement the pedestrian flow data typi-
cally gathered as part of a space syntax analysis with data on Blue-
tooth devices passing through pedestrian survey “gates.”4
However, approaches such as these can suffer from important limi-
tations: they rely on the deployment of ad hoc infrastructure or re-
3. R. Ahas and Ü. Mark, “Location Based Services—New Challenges for Plan-
ning and Public Administration?” Futures, vol. 37, no. 6, 2005, pp. 547–561.
4. E. O’Neill et al., “Instrumenting the City: Developing Methods for
Observing and Understanding the Digital Cityscape,” UbiComp 2006:
Ubiquitous Computing, LNCS 4206, Springer, 2006, pp. 315–332.

specificity of GPS, it nonetheless remains

attractive for urban research at scales
where this level of resolution is unnec-
essary. In effect, Erlang data provides
both a view of urban space as seen
through network bandwidth consump-
tion and, indirectly, insight into urban
life’s spatial and temporal dynamics.
This aspect makes it an excellent jump-
ing-off point for research supporting
public-transport planning, health and
safety, advertising, and other types of
group-directed activity.

First visualizations
and hypothesis
Figure 1 shows one of the simplest vi-
sualizations of Erlang data: a 3D plot
of telecommunications activity during
Madonna’s controversial 6 August
2006 performance, when more than
70,000 people converged on the Stadio
Olimpico for a concert condemned by
the Pope. Generic Erlang maps such as
this, which was presented at the Bien-
nale, are graphically appealing and
intuitively easy to grasp. However,
they’re actually quite difficult to inter-
pret rigorously, and they provide little
insight into local-area dynamics with-
out additional processing.
We hypothesize that by employing var-
Figure 1. A 3D plot of telecommunications activity during a Madonna concert
in Rome.
ious statistical techniques, we can use dif-
ferences in Erlang data over time to derive
clues to the types of activity in the imme-
diate area of the mast. (A mast can carry
multiple antennas, oriented in different
directions or serving different frequencies.)
This analysis is conceptually related to the
idea of a chronotype (see the “Chrono-
types and Space-Time Typologies” side-
bar), except that we’re characterizing
spaces by their mobile-bandwidth use over
time. By analyzing the bandwidth “signa-
ture” of each antenna, we try to envision
how it might correlate with urban activi-
ties in the geographical vicinity.
Because Erlang data is an antenna-
level measure, we needed an algorithm
to spread the point data values across the

JULY–SEPTEMBER 2007 PERVASIVE computing 31

 URBAN COMPUTING

Chronotypes and Space-Time Typologies

T o help conceptualize a city’s complex hourly, daily, weekly,

monthly, and annual rhythms, Luca Bertolini and Martin Dijst
put forward Roberta Bonfiglioli’s concept of the chronotype.1 The
typology of urban, suburban, and rural municipalities … based on
diurnal weekday variations in visitor populations”2 as a way to un-
derstand how place works in Manuel Castells’ “network society.”3
chronotype is a useful conceptual handle for thinking about how dif-
ferent groups occupy the same space depending on the time of day.
Bertolini and Dijst offer the example of a mixed-use area inhabited by REFERENCES
young couples without children and by families. The young couples 1. L. Bertolini and M. Dijst, “Mobility Environments and Network Cities,” J.
will likely work in another part of the city, returning perhaps only in Urban Design, vol. 8, no. 1, 2003, pp. 27–43.
the evening to socialize in bars and restaurants. In contrast, family
2. R. Zandvliet and M. Dijst, “Short-Term Dynamics in the Use of Places: A
members will go shopping and use other services during the day in Space-Time Typology of Visitor Populations in the Netherlands,” Urban
this area. The same space can thus have two or more distinct uses and Studies, vol. 43, no. 7, 2006.
populations.
3. M. Castells, “Grassrooting the Space of Flows,” Cities in the Telecommu-
As another way to conceptualize these rhythms, Robbert Zand- nications Age, J. Wheeler, Y. Ayoma, and B. Warf, eds., Routledge, 2000,
vliet and Dijst offer space-time typologies.2 That is, they propose “a pp. 18–27.

area served, accounting for distance
decay in signal coverage and multiple
antennas on a single mast. Carlo Ratti
and his colleagues took a center-of-
gravity approach,2 but to interpolate
values for the entire metropolitan region,
an alternative algorithm3 was used to
divide Rome into “pixels” measuring
1,600 m2. We used an exponential dis-
tribution function to derive an Erlang
point value based on a composite signal
from the surrounding masts.
We use this mathematical notation:
• Loc is the set of 1,600 m2 pixels.
• T96 is the set of times when we made
observations each day of the week.
Because we took measurements every
15 minutes, one day comprises 96
observations.
• Day is the set of {Weekday, Friday, Sat-
urday, Sunday} (we discuss this in
more detail later).
• erlang(␦, ␭, ) defines the Erlang value
at location ␭  Loc, at time   T96,
and ␦  Day.
{ }
• mean ai indicates the mean of the
i∈I
values ai, i  I.
(To preserve confidentiality, TI used a
scaling factor to adjust the Erlang values
transmitted to the SENSEable City Labo-
ratory. This means that while the rela-
tive difference between any two observa-
tions is scaled consistently, the actual
Erlang value at that point in time is un-
known. So, it’s helpful to focus on the
relationships between points over time
and space rather than the specific value
at any one point in time.)
Using prior knowledge of the city, we
arbitrarily selected eight locations that
we expected to have markedly different
signatures. Following an initial vi-
sualization exercise, we selected six for
analysis:
• Termini, Rome’s main passenger rail
station and busiest subway station;
• Trastevere, a mixed-use area popular
with Romans and tourists for its bars
and restaurants;
• the Piazza Bologna, a residential area
east of the city center;
• the area in front of the Pantheon (one
of Rome’s premier tourist attractions),
which also contains many bars and
restaurants;
• the Stadio Olimpico, a sports and ma-
jor concert venue northwest of central
Rome; and
• Tiburtina, a smaller rail and subway
interchange.
Figure 2 shows the pixels for these six
locations.
To minimize the impact of special
events on the data set, we calculated an
average Erlang value for each pixel at
each 15-minute interval, using a 90-day
period. So, for example, the data point
for 9 a.m. Monday is an average of every
9 a.m. Monday value between 1 Sep-
tember and 30 November 2006. We
excluded civic holidays from the calcu-
lation on the basis that they would intro-
duce unnecessary noise.
Erlang data by day of the week
Beginning with a minimal level of
processing, figure 3 shows how Erlang
data changes over time at each of the
six selected pixels. As the graphs in-
dicate, Monday through Friday are
broadly similar, except for a more rapid
decrease in activity on Friday after-
noon, suggesting a transition to the
weekend. Even more strikingly, Satur-
day and Sunday values often drop be-
low 50 percent of the typical weekday
load, but the drop’s magnitude varies
dramatically from site to site. This find-
ing indicates that weekday and week-
end data should be treated separately
in our analysis.
Intriguingly, areas more closely identi-

32 PERVASIVE computing www.computer.org/pervasive

Figure 2. A map of Rome indicating
the six locations (“pixels”) selected for
analysis of mobile phone usage.

fied with Roman residents, such as the

Piazza Bologna and Tiburtina, display
lower levels of day-to-day Erlang variance
than those associated with more transient
populations of commuters or tourists,
such as Termini and the Pantheon. This
peculiarity suggests that the greater the
flux of people—or, possibly, nonresi-
dents—through a site, the greater the vari-
ance in the signal. Conversely, predomi-
nantly “local” areas seem to have higher
levels of routine or habitual activity and
thus less variation between days.
However, in spite of the differences
between weekdays, Fridays, and the
weekend, figure 3 shows that all six loca-
tions demonstrate a broadly compara-
ble rhythm—a rapid ramping-up of
telecommunications activity between 6
and 10 a.m. on weekdays and a slower
pace on weekends. Apart from the Sta- 0 625 1,250 2,500 meters Selected pixels for analysis Subway stations
dio Olimpico, where the rhythms of con-
certs and football matches clearly show
on the graph, patterns are quite uniform:
a clear double peak and varying ratios
{ (
mean erlang δ , λ , τ
λ ∈Loc
)} bution of telecommunications activity.
Such events will likely place a corre-
of weekday-to-weekend activity. So, how spondingly high load on urban infra-
can we make differences in signatures 2. We then normalize the signature structure and resources, and a similar
between different sites more evident? over space: spike in Erlang data is also likely during
emergencies. This suggests that the real-
Normalization erlang norm _ space = time recognition of unusual concentra-
The magnitude of the differences be-
tween sites in figure 3 makes it hard to
(
erlang δ , λ , τ ) tions of telecommunications activity
might have relevance for public safety
compare them in a more detailed way, { (
mean erlang δ , λ , τ
λ ∈Loc
)} planning and transport scheduling.
so some type of data normalization is On weekends, the higher levels of
necessary. In figure 4, we plot the ratio of 3. We then normalize the signature over bandwidth use between midnight and 2
telecommunications intensity at one time: a.m. near the Pantheon and in Traste-
pixel against the average of every pixel in vere relative to work-oriented and resi-
the system at that point in time (nor- τ ∈T96 → erlang norm _ space _ time = dentially oriented pixels strongly suggest
malization over space). We then compare
that to the daily pixel average (normal-
(
erlang norm _ space δ , λ , τ ) leisure activity. This feature suggests that
we can also identify cultural and leisure
ization over time). Using this approach, {
mean erlang norm _ space δ , λ , τ
τ ∈T96
( )} areas on the basis of their telecommuni-
we can identify otherwise hidden shifts cations signature. Another feature with
in the relative intensity of activity across implications for the understanding of
Rome. The differences in figure 4 are quite urban dynamics is the high level of activ-
We employed these normalization steps: visible, and the radically different signa- ity at the transit hubs on weekday morn-
ture at the Stadio Olimpico indicates that ings, compared to residential sites.
1. For each location ␭ and day ␦, we you can readily recognize certain classes Figure 5 shows a more diffuse pattern
calculate of urban activity by the unusual distri- of spatial activity on weekends. This is

JULY–SEPTEMBER 2007 PERVASIVE computing 33

 URBAN COMPUTING
Stadio Olimpico
Erlang (natural scale)
Termini
Erlang (natural scale)
12 a.m. 5 a.m. 10 a.m. 3 p.m.
Monday
Figure 3. Erlang data for the six locations by day of week.
consistent with the idea that although
weekday telecommunications activity at
each site exhibits a more dynamic tempo-
ral pattern, weekend activity exhibits more
spatial dispersal. From an urban-planning
standpoint, this strongly suggests large
commuter flows into the central business
district during the week and more resi-
dentially oriented activity on weekends.
Of course, planners are well aware of this
spatial relationship, but spatial and tem-
poral visualization of these features at this
scale hasn’t been possible before.
One caveat: the levels of activity be-
tween 3 and 6 a.m. throughout the week
mean that any analysis using that period
would be rooted in extremely low Erlang
values. So, such a comparison might erro-
neously indicate excessive shifts in activ-
ity from site to site. Nonetheless, from this
initial analysis, it seems that through nor-
malized signatures we can reconstruct
some of the functioning of the city using
the invisible fingerprints of mobile phone
infrastructure.
34 PERVASIVE computing
Pantheon
Tiburtina
8 p.m. 12 a.m. 5 a.m. 10 a.m. 3 p.m. 8 p.m.
Tuesday Wednesday Thursday
Cluster analysis
So far, we’ve focused largely on indi-
vidual pixels, and we’ve identified some
interesting features at a fairly detailed
spatial level. Our preliminary analysis
indicates that residential areas, com-
muter hubs, nighttime hot spots, and
even special-event venues demonstrate
features consistent with our contextual,
anecdotal knowledge of Rome. How-
ever, validating our hypotheses requires
a more rigorously quantitative study.
The ultimate goal is to take the derived
signatures, group them by degree of sim-
ilarity, and map them to urban spa-
tiotemporal structures.
As a proof of concept, we created a
simplified vector—required for compu-
tational manageability—to feed pixel
data for each of Rome’s 262,144 pixels
to a clustering algorithm. An examina-
tion of our six selected pixels suggested
that six times in the daily cycle of Erlang
activity are particularly significant: 1
a.m., 7 a.m., 11 a.m., 2 p.m., 5 p.m., and
Piazza Bologna
Trastevere
12 a.m. 5 a.m. 10 a.m. 3 p.m. 8 p.m.
Friday Saturday Sunday
9 p.m. Each of these points lies toward
the middle of a period of rapid change
or significant variation between sites—
the early morning rise in activity, late
morning peak period, early afternoon
lull, afternoon peak, and evening drop.
The six normalized Erlang values thus
make up the coordinates of a vector that
describes, in a limited way, each pixel’s
signature.
We could use many clustering tech-
niques to create segmentations based on
the affinity between vectors. We chose a
K-Means approach, such that every
observation in a cluster is as much like
other members of that cluster and as dif-
ferent as possible from members of any
other cluster. With six coordinates from
each day, and separate sets of coordi-
nates for Monday through Thursday
(one set of averaged observations), Fri-
day, Saturday, and Sunday, the K-Means
algorithm used a 24-dimensional space.
We employed two clustering steps.
First, for each pixel, we calculate fea-
www.computer.org/pervasive
Figure 4. Erlang data normalized over
space and time by site: (a) Monday
4
through Thursday and (b) Saturday.

Normalized Erlang
3

2
ture(loc) = {erlang(␦, ␭, j)}, j = 1 a.m., 7
a.m., 11 a.m., 2 p.m., 5 p.m., and 9 p.m.
1
Second, the K-Means clustering algo-
rithm partitions the pixels into mutually 0
exclusive clusters. Each cluster is charac- 12 a.m. 5 a.m. 10 a.m. 3 p.m. 8 p.m.
terized by its centroid, and the algorithm (a)
aims to minimize the error function: Normalized Erlang 4

∑ distance (loc j , centroidk )

nClusters 2
∑
3

k =1 loc j ∈Clusterk
2
where clusterk is the set of objects related
to the cluster k, and centroidk is the mean 1
of all the points in clusterk. We calculated
the distance between pixels using the 0
12 a.m. 5 a.m. 10 a.m. 3 p.m. 8 p.m.
squared Euclidean distance: (b)
distance loc1 , loc2 = ( ) Termini Trastevere Pantheon
1
⎛ 24 Piazza Bologna Tiburtina Stadio Olimpico
2⎞
2

⎜∑
⎜⎝ τ =1 i
( )
feature loc1 − feature loc2 ( )i
As a result of the clustering process, we
can group all pixels in the city into any arbi-
trary number of groups based on the affin-
ity of their composite Erlang signature. In
our tests, we found a mix of clusters that
suggest a complex set of relationships
between signatures. Given the sheer com-
⎟
⎟⎠
plexity of cities, this is hardly surprising.
However, the existence of several small
clusters with much stronger levels of affil-
iation or differentiation indicates that the
overall data set includes some quite distinct
signatures. These signatures will likely map
to distinct types of urban activity.
For this initial research, we worked
with eight clusters as a compromise
between simplicity and specificity. Doing
this gave us a fair cophenetic correlation
value of 0.7704. Cophenetic correlation
is one way to gauge the clusters’ fit to the
original data set—values approaching
1.0 suggest a good fit—by comparing
pairwise linkages between observations.

Figure 5. Erlang data for Rome normalized over space and time. Intensities range from low (blue) to high (red).

1 a.m. 7 a.m. 11 a.m. 2 p.m. 5 p.m. 9 p.m.

Monday to Thursday
Saturday

JULY–SEPTEMBER 2007 PERVASIVE computing 35

 URBAN COMPUTING
(a)
Cluster 1 Cluster 2 Cluster 3
Figure 6. Analysis of eight clusters of Erlang data: (a) clusters 1–4; (b) a satellite view of Rome, for comparison; (c) clusters 5–8.
Projecting these clusters onto a map
of Rome (see figure 6) naturally indi-
cates that they’re closely linked to the
normalized Erlang signatures. The
edges of Rome’s urban core are clearly
visible, as are the hot spots of urban
activity straddling the Tiber River. The
map suggests an overall structure to the
city, with a correspondence between
levels of telecommunications activity
and types of human activity. At this
point, however, we can’t verifiably con-
nect cellular signatures to specific types
of human activity.
We then adjusted the metric to favor
the two most distinctive types of use
seen in the normalized graphs: early
morning use suggestive of commuting
behavior and late evening use suggestive
of nighttime leisure activities. For these
clusters we obtained cophenetic corre-
lations of 0.7630 and 0.8508, indicat-
ing that the clustering approach has sub-
stantial promise.
The red nighttime-leisure cluster in fig-
ure 7a shows two discrete spatial group-
ings that map anecdotally to known
areas of evening activity: Trastevere and
the area ranging to the west and south
of the Piazza Navona, and the vicinity
of the Piazza Spagna. The red commuter
clusters in figure 7b quite astonishingly
36 PERVASIVE computing
(b)
Cluster 4 Cluster 5
map to the most important points of
entry to the city by car and train: Ter-
mini station, Tiburtina, the end of the
Corso d’Italia, the Porta Maggiore, and
the Porta San Giovanni.
Discussion
Our preliminary findings suggest that
signature analysis can provide an impor-
tant new way of looking at the city as a
holistic, dynamic system. In particular, the
mobile phone network lets us develop a
real-time representation of those dynam-
ics at the city and city-region scale. This
approach can complement traditional col-
lection techniques, which are often out-
dated by the time they’re available to pol-
icy makers and the general public. Of
course, because our hypotheses so far are
based on anecdotal evidence, our findings
will require additional validation, which
we outline below.
What’s most promising about this
early research is the extent to which our
findings seem to parallel those of other
European researchers4,5 as well as more
conceptual research into telecommuni-
cations’ impacts on urban behaviors.6,7
In particular, we can characterize areas
on the basis of flows and dynamics
rather than on the basis of comparatively
static physical or demographic features.
(c)
Cluster 6 Cluster 7 Cluster 8
Moreover, we’ve recently received
data from Pagine Gialle (the Italian Yel-
low Pages) with which we intend to val-
idate our initial findings by linking the
signatures to spatial data on business
types and densities. In so doing, we can
build on the processing requirements we
discussed earlier in this article:
1. Antenna and pixel values must be
normalized over both space and
time to provide a measure of rela-
tive telecommunications intensity.
2. The substantial differences between
weekdays and weekends require
treating them separately in a classi-
fication algorithm.
3. The key time periods intimated in
this initial analysis appear to be 12
to 2 a.m., 5 to 8 a.m., 10 a.m. to 12
p.m., 2 to 6 p.m., and 8 to 11 p.m.
However, as our initial cluster analy-
sis makes clear, these aren’t the only
factors.
We expect several other analytical
approaches to yield insights into net-
work usage patterns. One of the most
promising approaches is Eigenbehavior
analysis.8 Because we can easily map the
signature to a vector representation of
the sort already used in the cluster analy-
www.computer.org/pervasive
Figure 7. Analysis of the five clusters
covering Erlang data for the two most
distinctive types of cell phone use:
(a) nighttime leisure, (b) early morning
commuting.
sis, deriving the Eigenvectors should be
quite straightforward. Applying other
analytical techniques such as Fourier and
wavelet transform plots might reveal
new, distinct characteristics. Ideally, each
of these analyses will eventually feed into
a single categorization process that can
discriminate between discrete types of
behavior at the antenna and pixel levels.
Limitations of research
As we mentioned before, TI’s masking
function meant that we weren’t able to
work with true Erlang values. An addi-
tional constraint is that owing to opera-
tional requirements and planning re-
strictions, GSM masts are irregularly
distributed and oriented. To manage the
computational and mathematical com-
plexity of calculating point values for
262,144 pixels over a three-month per-
iod, our algorithm spreads Erlang data
through all 360 degrees, producing a pos-
sible skew in the overall distribution.
Finally, not all masts handle both the
900- and 1,800-Hz bands used in
Europe. So, some network activity might
gravitate toward more physically remote
base stations with the hardware to
process calls in a particular band. We
don’t have data that would let us com-
pensate for these possible biases. So,
without adopting an entirely different
approach to data collection—one that
the network operator would have been
reluctant to support at this development
stage—localizing phones more accu-
rately is impossible.
Although the data to which we cur-
rently have access has clear, substantial
limitations, we believe our approach rep-
resents an appropriate trade-off between
locational specificity and implementa-
tional feasibility. Fortunately, analysis at
the city and city-regional scale doesn’t
depend on the high level of accuracy that
JULY–SEPTEMBER 2007
Nighttime-leisure-weighted clusters
(a)
Cluster 1 Cluster 2
location-based services typically require.
So, we feel that there’s plenty of oppor-
tunity to gain valuable insights into
urban dynamics using Erlang data at
smaller scales, and we intend to move
forward with it.
I
t would be exciting to compare the
signatures collected from Rome
with similar data from other major
European cities such as London,
Paris, or Frankfurt. For instance, it’s rea-
sonable to expect that cities with more
distinct spatial patterns of human activ-
ity might display correspondingly more
distinct patterns of network use and
more readily classifiable signatures.
Unfortunately, at this time commercial
considerations appear to preclude using
data from other network operators.
This issue highlights the extent to
which research using cellular networks
must take nonscientific factors into
account. First, a policy framework at the
national or European level that encour-
ages networks to share nonidentifiable
data with planning and policy researchers
would be immensely helpful. Clearly,
there are important considerations from
the standpoint of commercial confiden-
tiality, personal privacy, and possibly
even national security. However, in the
absence of clear regulatory guidance, fur-
Transit-weighted clusters
(b)
Cluster 3 Cluster 4 Cluster 5
ther research using cell phones—the most
widely deployed device with locational
capabilities—won’t be possible at the city
or city-region scale.
Without encouragement, far more
detailed data sets held by the networks
will never see the light of day. For in-
stance, paging data—generated by
polling the phones in a cell to obtain a
list of IMEI (International Mobile Equip-
ment Identity) numbers at the mast
level—could provide unmatched detail
on travel origins and destinations, and
on population densities. By scrambling
handset identifiers with changing encryp-
tion schemes, reporting only partial tra-
jectories, and never reporting on cells or
paths containing fewer than an agreed
minimum number of users, you would
be able to perform this kind of research
without compromising personal privacy.
This data would also assist enormously
in understanding how individual and
group behavior changes over time and
space. This would not only shed further
light on the rhythms of urban life but also
address the fact that you can’t derive met-
rics on activity and population densities
from Erlang data alone.
The challenge is that as the data
becomes more useful, it also becomes
more sensitive to both operators and end
users. An all-or-nothing approach to pri-
vacy has hampered this discussion.
PERVASIVE computing 37
 URBAN COMPUTING

the AUTHORS
Jonathan Reades is an MPhil and a PhD candidate at the Bartlett School of Planning
at University College London. His research interests are the application of mobile
phone data to topics in urban planning such as business clustering and communica-
tion, and the spatiotemporal structures of European cities. He previously spent eight
years in data analytics, helping telecom firms use their data for targeted marketing. REFERENCES
His first degree was in comparative literature at Princeton University; he’s a student
member of both the Royal Town Planning Institute and the Town and Country Plan- 1. F. Calabrese and C. Ratti, “Real Time
ning Association. Contact him at the Planning Dept., 4th fl., The Bartlett School, Rome,” Networks and Communication
Wates House, 22 Gordon St., London WC1H 0QB, UK; j.reades@ucl.ac.uk. Studies, vol. 20, nos. 3 & 4, 2006, pp.
247–258.

2. C. Ratti et al., “Mobile Landscapes: Using

Francesco Calabrese is pursuing a PhD in informatics and automatics engineering Location Data from Cell-Phones for Urban
from the University of Naples Federico II and is a visiting doctoral student at the Mass- Analysis,” Environment and Planning B:
achusetts Institute of Technology’s SENSEable City Laboratory. His research interests Planning and Design, vol. 33, no. 5, 2006,
include hybrid control systems, embedded control systems, computer–numerical- pp. 727–748.
control machines, finite-time stability, and real-time monitoring and analysis of urban
dynamics. He received his Laurea in informatics engineering from the University of 3. F. Calabrese et al., “Real-Time Urban
Naples Federico II. He’s a member of the IEEE and the IEEE Control Systems Society. Monitoring Using Cellphones: A Case-
Contact him at the SENSEable City Laboratory, MIT 10-485, 77 Massachusetts Ave., Study in Rome,” SENSEable Working
Cambridge, MA, 02139; fcalabre@mit.edu. Paper, SENSEable City Laboratory, Mass-
achusetts Inst. of Technology, 2007;
http://senseable.mit.edu/papers/pdf/
Andres Sevtsuk is a PhD candidate in city design and development and urban infor- CalabreseRatti2007SCLWorkingPaper.pdf.
mation systems at the Massachusetts Institute of Technology. His research interests
include accuracy and reliability in using mobile phone data for estimating the distri- 4. L. Bertolini and M. Dijst, “Mobility Envi-
bution of people in cities, the relationship between urban form and movement, and ronments and Network Cities,” J. Urban
the study of new possibilities in street parking and one-way vehicle rentals using dis- Design, vol. 8, no. 1, 2003, pp. 27–43.
tributed computation and sensing. He received his MS in architecture studies from
MIT. Contact him at 7 Dodge St., Cambridge, MA 02139; asevtsuk@mit.edu. 5. R. Zandvliet and M. Dijst, “Short-Term
Dynamics in the Use of Places: A Space-
Time Typology of Visitor Populations in the
Netherlands,” Urban Studies, vol. 43, no. 7,
Carlo Ratti is an associate professor of the practice of urban technologies at the Mass- 2006.
achusetts Institute of Technology, where he directs the SENSEable City Laboratory. He’s
also founding partner and director of carlorattiassociati, an architectural firm. He re- 6. S. Graham, “Cities in the Real-Time Age:
ceived his PhD in architecture from the University of Cambridge. He’s a member of the The Paradigm Challenge of Telecommuni-
Ordine degli Ingegneri di Torino and the Association des Anciens Elèves de l’École Na- cations to the Conception and Planning of
tionale des Ponts et Chaussées, and is a UK Registered Architect. Contact him at the Urban Space,” Environment and Planning
SENSEable City Laboratory, MIT 10-485, 77 Massachusetts Ave., Cambridge, MA 02139; A, vol. 29, no. 1, 1997, pp. 105–127.
ratti@media.mit.edu.
7. M. Sheller and J. Urry, “The New Mobili-
ties Paradigm,” Environment and Planning
A, vol. 38, no. 2, 2006, pp. 207–226.

8. N. Eagle and A. Pentland, “Eigenbehaviors:

It would be helpful to move toward a
more nuanced understanding of how to
preserve reasonable expectations of pri-
vacy by the network user while creating
mechanisms to permit future research.
We need to establish the extent to which
certain types of data and analysis create
either the perception of a privacy inva-
sion9 or the real risk of trail reidentifi-
cation,10 and to set out the trade-offs for
public review and discussion.
ACKNOWLEDGMENTS
The Real Time Rome exhibit at the 10th Interna-
tional Architecture Exhibition at Venice, Italy was
sponsored by Telecom Italia, with technical part-
ners Biennale di Venezia, the City of Rome, Google,
Atac-Rome, Samarcanda, and Microtek. We grate-
fully acknowledge the significant contributions of
Burak Arikan, Assaf Biderman, Filippo Dal Fiore, Sa-
ba Ghole, Daniel Gutierrez, Sonya Huang, Sriram
Krishnan, Justin Moe, Francisca Rojas, and Najib
Marc Terazi.
We also thank Matthew Jull, Tim Kindberg, and our
anonymous reviewers for their valuable feedback;
Giovanni Celentano of the University of Naples and
Lucio Pinto of the Silvio Tronchetti Provera Founda-
tion for their generous suggestions; and Peter Hall
and Michael Batty of University College London and
Larry Vale of the Massachusetts Institute of Techno-
logy for fostering a collaborative research environ-
ment. And a final thank-you to the Balzan Founda-
tion, whose support for young researchers through
the Balzan Prize helped enable this transatlantic
collaboration.
Identifying Structure in Routine,” 2006;
http://vismod.media.mit.edu/tech-reports/
TR-601.pdf.
9. M. Langheinrich, “Privacy Invasions in
Ubiquitous Computing,” 2002; http://
guir.berkeley.edu/pubs/ubicomp2002/
privacyworkshop/papers/uc2002-pws.pdf.
10. B. Malin, “Betrayed by My Shadow: Learn-
ing Data Identity via Trail Matching,” J.
Privacy Technology, 2005, pp. 1–22; www.
jopt.org/publications/20050609001_malin
_abstract.html.
For more information on this or any other comput-
ing topic, please visit our Digital Library at www.
computer.org/publications/dlib.

38 PERVASIVE computing www.computer.org/pervasive